AI Agents and Agentic AI in Medicine: An Autonomy-Centred Scoping Review and Research Agenda

Abstract Artificial intelligence agents—software systems that perceive, reason, plan and act over multiple steps with limited human intervention—are moving from research prototypes towards clinical deployment. Recent reviews map what has been built; this one asks what would have to hold for such systems to be deployed safely. We conducted a scoping review reported according to the PRISMA extension for scoping reviews, searching eight bibliographic sources from January 2017 to June 2026, admitting foundational computer science and regulatory literature alongside clinical applications. Five architectural paradigms recur—tool-augmented large language model agents, multi-agent systems, offline reinforcement learning agents, embodied robotic agents and autonomous monitoring agents—nearly all sharing a common four-layer stack. We introduce a six-level scale of clinical agent autonomy that grades systems by where human judgement enters relative to the reasoning, and by whether the resulting action remains reversible. The scale exposes an inversion in the evidence base: demonstrated patient benefit concentrates at high autonomy in narrow, bounded domains, whereas research effort concentrates at intermediate autonomy, where prospective outcome evidence is absent and automation bias is the dominant risk. We further propose seven evaluation dimensions specific to agency, a mapping of evaluation requirements to autonomy level, and a ten-item failure mode taxonomy assigned to the architectural layer at which each is addressable. Deploying agentic systems safely requires prospective multi-site validation, reporting standards adapted to agency, equity-aware design, and regulatory obligations indexed to autonomy rather than algorithm class. We conclude with an eight-point research agenda.